Reliability lives
in the edges.
Robo Co-op's development skill system for Claude Code. Lite, Standard, and Pro share one conviction: never let the agent that wrote the code decide whether it's good enough. Reading the prompt won't find the bugs — running it will.
/plugin marketplace add Robo-Co-op/robobuilder-standard /plugin install robobuilder@robo-coop-tools /reload-plugins /robobuilder:start
robobuilder isn't a skill library.
It's an argument.
An argument about how AI-assisted engineering should verify itself. Four convictions — each one earned by a real incident in this system's own history, not a design meeting.
Never let the maker grade its own work.
tdd-pair detects, as an observable fact, any cycle that produced code
without a new failing test first, and returns nothing but OK or VIOLATION.
cross-review runs independent agents in parallel. Pro's
champion-challenger applies the same rule one level up — to a loop's own
attempted improvements. "Use this when you notice yourself cutting corners" is
rejected on principle: it hands the decision back to the exact judgment it exists
to replace.
"Never measured" was quietly counted as "passed."
The most common defect class in this system's history. guard advertised
"full safety mode" while its three PreToolUse hooks pointed at directories that exist
in no edition — enforcing nothing at all. learn compared a field no
writer ever emitted, so NaN > NaN was always false, and reported an
empty log as a clean result. To a system, absent measurement and a clean pass look
identical. Telling them apart is now a guard's whole job.
Prose review doesn't find it. Execution does.
An agent-team sweep ran every skill that had never once been executed. 114 suspected defects came back; independent verifiers refuted 63 of them; the rest were confirmed and fixed. Eleven-plus rounds of prompt-only review, run before that sweep, had found none of them.
Put a graph outside the loop.
Standard covers the inner loop — L1 (the agent loop) and L2 (verification: tdd /
diagnose / review). Pro adds the outer loop — L3 (event-driven) and L4
(self-improving) — and above it L5: a layer answering, with topology,
four failure modes no single loop can reach from the inside. L5 is
robobuilder-pro's own extension; it is never credited to the swyx / LangChain L1–L4
scale it cites. Cite the source, never claim its credit.
| Failure mode | Why the loop can't see it from inside | The topological answer |
|---|---|---|
| Metric gaming | It is succeeding by its own measure — nothing inside disagrees | Pairing — a watching loop on a counter-metric |
| Upward blindness | The target it chases is an input, not something it may question | Hierarchy — a slower loop owns that target |
| Inter-loop conflict | Neither loop can observe the other's effect | Arbitration — a loop above owns the trade-off |
| Measurement decay | It trusts its own inputs by construction | Audit — an independent loop anchored to ground truth |
Three editions.
One workflow.
Learn it on Lite, run it daily on Standard, add Pro when you need it — one design at three concentrations.
| Edition | Who it's for | What's inside |
|---|---|---|
| Lite v1.3.0 | Beginners — learn the workflow in four commands | plan / build / improve / ship, plus the same hooks and review agents |
| Standard v1.8.0 | Daily development | 41 skills across 6 phases, 9 subagents, 6 lifecycle hooks, 3 playbooks |
| Pro v::error::robobuilder-pro returned an unparseable version: '' | Loop & Graph Engineering — autonomous agent loops | An add-on beside Standard: design, gate and audit loops, then wire them into a graph with typed edges |
/plugin marketplace add Robo-Co-op/robobuilder-standard /plugin install robobuilder@robo-coop-tools /reload-plugins /robobuilder:start // points you to your first 3 skills /robobuilder:tune-claude-md // personalizes your CLAUDE.md
$ /robobuilder:build "fix the auth bug"
Pick → shows the ready issues → you choose → moves it to in-progress
Bug branch → builds a reproduction loop → five ranked hypotheses, one variable at a time
Fix → regression test →
tdd-pairconfirms RED came before GREEN$ /robobuilder:ship
pre-flight → CHANGELOG drafted → PR opened → waits for CI to go green → checks production
A ledger of bugs found
by running the code.
Every version below records what execution caught as much as what shipped. Eleven-plus rounds of prompt-only review, run before any of it, caught none of these.
| v1.5.1 | fixed | Reachability — every continue to Step N jump checked against skipping a mandatory section |
| v1.6.0 | fixed | Agent wiring — five of nine bundled agents were dispatched by no skill at all. A naive string match found only three; two more were hiding inside a worked example's console output |
| v1.7.0 | fixed | The runtime contract itself — eleven skills, 124 call sites, invoked the helper binaries by a bare relative path that never resolves from an installed plugin. Every read came back silently empty; every write silently failed |
| v1.8.0 | fixed | Closed all 51 confirmed findings from the execution sweep — including guard, which enforced nothing at all |
| v1.2.0 | added | The L5 graph layer — graph-design, champion-challenger, graph-audit |
| v1.2.1 | fixed | The same unwired-agent defect class Standard found, closed and pinned by a test |
| v1.1.0 | added | plan and build rebuilt to serve existing code, not only greenfield features |
| v1.1.1 | fixed | Two of six bundled agents were unwired — tdd-pair is the maker-≠-checker safeguard itself |
| v1.2.0 | added | A defence — a guard, hook, validator, auth rule — now forces improve --deep with 20+ bypass attempts built independently of the implementation. A shipped hook with 37 green tests and three reviewers still fell to 13 of 26 independently-built bypasses: the suite and the code shared one blind spot |
Every number on this page is under test.
A landing page is where a stale number survives longest — nothing breaks and nobody
notices. So each figure here is declared once in site/claims.json with the
source that derives it, and test_site_claims.py re-derives all of them
from the repository on every CI run. A claim with no source fails the suite; so does a
figure that appears in one language and not the other. Publishing the page and proving
it are the same commit.